Dolphins in the Net: Internet Content Filters and the Advocate General's Glawischnig-Piesczek v. Facebook Ireland Opinion
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
STANFORD
Dolphins in the Net:
Internet Content Filters and the
Advocate General’s
Glawischnig-Piesczek v.
Facebook Ireland Opinion
Daphne Keller
Director of Intermediary Liability
2019
559 Nathan Abbot Way Stanford, CA 94305 cyberlaw.stanford.eduDolphins in the Net:
Internet Content Filters and the Advocate General’s Glawischnig-
Piesczek v. Facebook Ireland Opinion
Daphne Keller 1
Stanford Center for Internet and Society
September 4, 2019
Table of Contents
I. Introduction...................................................................................................... 2
II. Process and Representation Issues .................................................................. 7
III. Making Facebook Monitor and Filter Users’ Posts ......................................... 11
A. Policy Backdrop ........................................................................................... 12
B. Legal Analysis .............................................................................................. 15
1. Fundamental Rights .....................................................................................17
a. What “Identical” or “Equivalent” Content Is to Be Filtered? .................. 19
i. “Identical” Content ................................................................................ 19
ii. “Equivalent” Content............................................................................ 22
b. How Much of the Filtered Content Will Be Illegal?................................. 23
c. What Happens if Filters Take Down Legal Content? ............................... 26
2. The eCommerce Directive ........................................................................... 28
a. Can Courts Order Platforms to Filter All User Content under
Article 15? ................................................................................................. 29
b. Does Filtering Cause Hosts to Lose Immunity under Article 14? ........... 31
c. Does Human Review Cause Hosts to Lose Immunity under Article 14? . 33
IV. Allowing Austria to Order Global Content Takedowns.................................. 35
V. Conclusion ...................................................................................................... 39
1 Directorof Intermediary Liability, Stanford Center for Internet and Society (Stanford CIS);
former Associate General Counsel to Google Inc. Stanford CIS’s funding and other information is
at http://cyberlaw.stanford.edu/about-us.
1I. Introduction
One of the most important Internet law cases in recent years, Glawischnig-
Piesczek v. Facebook Ireland, is currently pending before the Court of Justice of
the EU (CJEU). The case, which concerns a Facebook user’s vulgar comments
about an Austrian politician, has received surprisingly little attention. The
Advocate General (AG) issued his influential Opinion in June, and his
recommendation for the Court’s final ruling is troubling.2 It would open the door
to court orders requiring platforms to automatically detect and filter out
particular content posted by users. It would also let national courts order
platforms to globally remove users’ posts, even in countries where the posts
would be legally protected expression or information. The AG’s analysis has
serious implications for the EU’s Intermediary Liability legal framework as well
as currently-pending legislative changes. We should hope the Court does not
adopt it.
As the European Commission pointed out at the hearing, the Court’s Judgment
will have consequences far beyond Facebook or this specific case. It is likely to
shape the behavior of both large and small Internet platforms for years to come.
By doing so, it will indirectly but seriously affect Internet users’ rights, including
rights to privacy and freedom of expression and information. The Court’s
conclusions will also likely influence the choices EU lawmakers make as they
revise the EU’s primary Intermediary Liability law, the eCommerce Directive.
The dispute in this case started when a Facebook user posted a news article about
Eva Glawischnig-Piesczek, who then headed Austria’s Green Party. The user’s
comments next to the article criticized the party’s policy on refugees, and called
its leader a “lousy traitor” (miese Volksverräterin), a “corrupt oaf” (korrupter
2 Opinion of Advocate General Szpunar, Case C-18/18, Eva Glawischnig-Piesczek v. Facebook
Ireland Limited, (June 4, 2019)
http://curia.europa.eu/juris/document/document.jsf?text=&docid=214686&pageIndex=0&docla
ng=en&mode=req&dir=&occ=first&part=1&cid=7255506.
2Trampel), and a member of a “fascist party” (Faschistenpartei).3 Vulgar terms
like that in a political context would be protected expression in many countries.
But Glawischnig-Piesczek told Facebook they constituted defamation in Austria.
Facebook took the position that the comments were not clearly unlawful.
Therefore, it argued, under Austrian law, a court rather than a private company
should decide whether the post should come down—and whether Glawischnig-
Piesczek’s reputational rights outweighed the user’s expression rights and other
users’ rights to access information.
Glawischnig-Piesczek initiated two proceedings. In one, a criminal case, the court
ultimately said that the post was not clearly unlawful. In the second, a civil case, a
different court said that the post was obviously unlawful, and that Facebook was
therefore liable for failing to remove it after receiving notice. The civil case is the
one on appeal here. In an expedited preliminary injunction proceeding, the first
instance court ordered Facebook to proactively monitor and block “identical” and
“equivalent” posts in the future. On appeal, a higher court upheld a monitoring
requirement for “identical” posts. Both parties then appealed to the Austrian
Supreme Court. The appeal raised both the filtering issue and an issue not
addressed in the lower courts: whether Facebook must remove the post globally,
even in countries where it is legal.4 The Austrian Supreme Court referred both the
question about proactive monitoring and the question about global removal to
the CJEU. In his Opinion, the AG advises the Court to rule that both are
permissible in certain circumstances.
In this White Paper, I identify problems with the AG’s recommendations,
focusing in particular on the issue of filters. Many of the problems stem from the
rushed process by which the case reached the CJEU, and the limited voices and
3 The full post, informally translated, read “Lousy traitor. This corrupt oaf has never earned a
single honest cent with real work in her whole life, but with our tax money is kissing the asses of
these smuggled-in invaders to build them up into the most valuable of all. Let us finally prohibit
this Green Fascist party.”
4 The Austrian appellate court analyzed whether Austrian law applied to the case, but not what
geographic scope the injunction should have.
3perspectives represented in the litigation. I will review this procedural
shortcoming in Section II. The limited briefing on important aspects of the
filtering issue seems to have affected the AG’s conclusions about both
fundamental rights and the eCommerce Directive. In Section III, I will analyze
those legal conclusions, and suggest that they are inconsistent with the
requirements of the EU Charter and the Court’s precedent. In particular, they do
not adequately address the problem of “dolphins in the net” – lawful user
communications that are accidentally blocked by filters. Finally, in Section IV, I
will very briefly discuss the issue of global content removal.
The AG in Glawischnig-Piesczek, Maciej Szpunar, is a capable jurist who has
worked on important related cases. That includes two pending disputes about
Google’s implementation of the “Right to Be Forgotten,” one of which raises a
question analogous to the one in this case about global enforcement.5 For the
filtering questions in the Facebook case, unfortunately, his analysis of core issues
is opaque. He says courts can compel Facebook to filter “identical” information
across the entire platform, as well as “equivalent” information posted by the same
user. But it is not clear what information he considers “identical” or “equivalent.”
Nor is it apparent what portion of the material blocked by the proposed filter
would actually violate the law, or what corrective measures would be available if a
filter suppressed the wrong expression and information.
That imprecision will make it very hard for the Court to assess the proposed
filtering injunction’s consequences for fundamental rights. Neither the AG nor
the Court has the information needed to meaningfully weigh the injunction’s
benefits for the plaintiff against the burdens it will create for Facebook’s other
users. That includes potential harm to those users’ data protection rights and
5Opinion of AG Szpunar, Case C-507/17, Google v. Commission nationale de l’informatique et
des libertés (CNIL), (January 10, 2019) (“Google Op.”)
http://curia.europa.eu/juris/document/document.jsf?text=&docid=209688&pageIndex=0&docl
ang=EN&mode=req&dir=&occ=first&part=1&cid=9734778. Groups representing journalists and
other affected entities were able to intervene before the lower court in the Google case, giving the
AG and Court the benefit of additional input in that case.
4expression and information rights, as the Court has recognized in the past. Their
rights to equality and non-discrimination are also implicated: recent research
suggests that filtering errors disproportionately affect lawful speech by members
of racial and linguistic minority groups.6 Overall, Facebook’s users number in the
billions, so mistakes affecting even a tiny percent will still be very consequential.7
The AG and Court are not alone in lacking key information. Technologists and
civil society groups engaged in the EU’s ongoing policy debate about filters have
charged that the public overall has inadequate information about their real
capabilities and error rates. A 2019 letter from two dozen civil society
organizations to the EU Parliament said that even some of the most widely
deployed filters are “untested and poorly understood technologies,” with great
potential to harm “democratic values and individual human rights.”8
Still, the public debate has generated some relevant information, little of which
seems to have been surfaced to the court in this case. Civil society groups,
technologists, and human rights bodies have all raised concerns about states
relying on privately operated software to restrict expression and information.9
European legal experts have carried out detailed analyses of hosts’ immunities
and filtering proposals under the eCommerce Directive and fundamental rights
6 See Sap et al (2019) and Center for Democracy and Technology (2017), below note 11.
7 Facebook reports 2.41 billion monthly active users. Company Information,
https://newsroom.fb.com/company-info/.
8 Letter to Members of European Parliament (February 8, 2019)
https://cdt.org/files/2019/02/Civil-Society-Letter-to-European-Parliament-on-Terrorism-
Database.pdf. The letter continued,
The European public is being asked to rely on claims by platforms or vendors
about the efficacy of [a widely used filtering] Database and similar tools—or
else to assume that any current problems will be solved by hypothetical
future technologies or untested, post-removal appeal mechanisms. Such
optimistic assumptions cannot be justified given the serious problems
researchers have found with the few filtering tools available for independent
review. Requiring all platforms to use black-box tools like the Database
would be a gamble with European Internet users’ rights to privacy and data
protection, freedom of expression and information, and non-discrimination
and equality before the law. That gamble is neither necessary nor
proportionate as an exercise of state power.
9 See notes 10 through 15.
5law.10 This information is commonplace in discussions in Brussels, but, because
of limits on who gets to participate in the case, does not seem to be before the
Court.
A poorly considered ruling in this case could have major unintended
consequences for the fundamental rights of Internet users around the world. To
avoid those, the Court should take care, in its ultimate ruling, to articulate very
clearly what questions it is resolving—and what questions it is not. It should not
rely on untested claims or assumptions about filtering technology, or resolve legal
questions that were not squarely raised and adequately briefed. It should, I
believe, reject the Austrian court’s injunction. If it does implicitly or explicitly
endorse filtering in some circumstances, it should be very clear about how
filtering injunctions can be reconciled with users’ fundamental rights. At a
minimum, injunctions should not issue without very strong showings that filters
provide the best means to protect a plaintiff’s or government’s interests, and will
not generate mistakes and false positives that disproportionately burden the
rights of other Internet users. To the extent that the Court’s conclusions require
further fact-finding or legal analysis, its Judgment should make that clear.
Clarity in the Court’s Judgment will be essential to the Austrian courts as this
case continues. So will the Court’s guidance about fundamental rights and the
eCommerce Directive. EU policymakers, too, will look to the Court’s Judgment in
order to understand the rights and interests at stake as they revise the EU’s
platform liability laws.
10See, e.g., Christina Angelopoulos, On Online Platforms and the Commission's New Proposal
for a Directive on Copyright in the Digital Single Market, (2017),
https://ssrn.com/abstract=2947800; Sophie Stalla-Bourdillon, Eleonora Rosati, Karmen
Turk, Christina Angelopoulos, Aleksandra Kuczerawy, Miquel Peguera and Martin Husovec, A
Brief Exegesis of the Proposed Copyright Directive, (2017),
https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2875296; Joris van Hoboken, João Pedro
Quintais, Joost Poort, and Nico van Eijk, Hosting intermediary services and illegal content
online: An analysis of the scope of article 14 ECD in light of developments in the online service
landscape, (2019),
https://www.ivir.nl/publicaties/download/hosting_intermediary_services.pdf.
6II. Process and Representation Issues
The procedure in this case has effectively deprived both the AG and Court of
crucial information and analysis. Most importantly, they are not hearing from
some of the people who will be most affected by the ruling, and are not hearing
information and arguments that are familiar to participants in the public policy
debate. That’s troubling both for the likely quality of the resulting Judgment, and
for its legitimacy in the eyes of European policy stakeholders.
Three factors in particular contribute to this problem. First, the case arose from
an expedited preliminary injunction proceeding, so the factual and legal record
before the Court is very slight. The Austrian courts never even addressed the
complex questions around global takedown, for example. On filtering issues, the
appeals court’s discussion focused almost entirely on defamation law and harm to
the plaintiff. It did not analyze the efficacy of filters in redressing that harm, or
the unintended consequences the filters might have for other Facebook users.
Second, the advocacy in this case is imbalanced. As is common in Intermediary
Liability cases, the people most concerned about expression and information
rights are not parties, so the courts never hear from them. The plaintiff is the
person who was harmed by an online post, but the defendant is not the post’s
author. Instead, the defendant is Facebook—a technology company, with
interests that inevitably may not align with those of its users or other people in
the Internet’s information ecosystem.
Both of those problems might be addressed by intervention and briefing from
civil society organizations or other expert groups. But those organizations did not
recognize the case’s importance in time to petition national courts to intervene in
the expedited Austrian proceeding, and were not able to weigh in once the case
arrived at the CJEU. That lack of input is the third, and perhaps most serious,
factor distorting the picture presented to the Court and the AG. The only
perspectives they see are those of the plaintiff, Facebook, and five government
7interveners (Austria, Latvia, Portugal, Finland, and the EU Commission). The
government briefs are confidential, so we don’t know what they said. But it seems
unlikely that they provided insight into evolving technology, which is at the heart
of this case. And while the briefs from Member States likely represent the public
interest as they see it, government lawyers may not be the best advocates for
expression and information rights in a case about vulgar criticism of a politician.
The filtering debate as presented in the case looks very different from the one
that participants in Brussels and throughout the EU would recognize. Many
technologists,11 academics,12 and civil society organizations like EDRi,13 Article
19,14 and Access Now15 have argued that filters pose major and poorly understood
threats to fundamental rights. Human rights bodies, too, have raised serious
concerns.16 Facebook, in that larger public debate, is somewhere in the middle. It
supports filters in some circumstances, and has come under sustained criticism
for doing so. (To be clear, I have never heard of Facebook representatives
endorsing a filter like the one proposed in this case, which would search users’
textual posts for defamation—indeed, I have heard them express great skepticism
11 See, e.g., Evan Engstrom and Nick Feamster, The Limits of Filtering: A Look at the
Functionality and Shortcomings of Content Detection Tools, (2017), www.engine .is/the-limits-
of-filtering; Center for Democracy and Technology, Mixed Messages? (2017),
https://cdt.org/files/2017/11/Mixed-Messages-Paper.pdf; Maarten Sap et al, The Risk of Racial
Bias in Hate Speech Detection, (2019),
https://homes.cs.washington.edu/~msap/pdfs/sap2019risk.pdf.
12 See, e.g., Angelopolous, supra note 10.
13 European Digital Rights, Press Release: Censorship machine takes over EU’s internet, (March
26, 2019) https://edri.org/censorship-machine-takes-over-eu-internet/.
14 Article 19, Facebook congressional testimony: “AI tools” are not the panacea, (April 13, 2018)
https://www.article19.org/resources/facebook-congressional-testimony-ai-tools-not-panacea/.
15 Denis Nolasco and Peter Micek, Access Now responds to Special Rapporteur Kaye on “Content
Regulation in the Digital Age”, AccessNow (January 11, 2018)
https://www.accessnow.org/access-now-responds-special-rapporteur-kaye-content-regulation-
digital-age/.
16 David Kaye, Joseph Cannataci and Fionnuala Ní Aoláin, Mandates of the Special Rapporteur
on the promotion and protection of the right to freedom of opinion and expression; the Special
Rapporteur on the right to privacy and the Special Rapporteur on the promotion and protection
of human rights and fundamental freedoms while countering terrorism, (December 7, 2018),
https://spcommreports.ohchr.org/TMResultsBase/DownLoadPublicCommunicationFile?gId=24
234; Council of Europe, Recommendation CM/Rec(2018)2 of the Committee of Ministers to
member States on the roles and responsibilities of internet intermediaries, Committee of
Ministers (March 7, 2018),
https://search.coe.int/cm/Pages/result_details.aspx?ObjectID=0900001680790e14.
8about filtering text generally.) Facebook is a major proponent and backer of a
controversial filter17 for violent extremist videos and images, and its founder has
extolled18 the potential for AI-based content moderation generally.
In the circumscribed world of this case, though, Facebook is the sole filtering
skeptic. All the other briefs apparently maintain that using software to
automatically block defamation of a public figure while still respecting users’
fundamental rights “must be possible.”19 If public interest organizations and
other interested groups had submitted briefs in the case, the Court would have
heard strong criticisms of that idea. In this essay, I will lay out many of those
criticisms as they relate to expression and information rights. But there are other
relevant areas where the interests of platforms and those of users affected by
filters diverge. The lack of adequate briefing on those is a problem, too. By way of
illustration, these include:
Privacy and Data Protection: The Court has clearly said in the past that using
filters to automatically scan and assess users’ every communication can burden
their data protection rights. This issue barely comes up in the AG’s Opinion,
though. Had civil society groups intervened, that issue—and deeply intertwined
questions about the monitoring that Facebook already uses to target ads—would
almost certainly have been fleshed out. Privacy experts could have weighed in, for
example, on the relevance of users’ rights, under the GDPR, in relation to the
kinds of automated decision-making carried out by filters. They could have
discussed CJEU cases like Digital Rights Ireland, which rejected a law requiring
electronic communications service providers to retain data about
17 Olivia Solom, Facebook, Twitter, Google and Microsoft team up to tackle extremist content,
The Guardian (December 6, 2016)
https://www.theguardian.com/technology/2016/dec/05/facebook-twitter-google-microsoft-
terrorist-extremist-content.
18 Sydney Li and Jamie Williams, Despite What Zuckerberg’s Testimony May Imply, AI Cannot
Save Us, Electronic Frontier Foundation (April 11, 2018)
https://www.eff.org/deeplinks/2018/04/despite-what-zuckerbergs-testimony-may-imply-ai-
cannot-save-us.
19 Par 55.
9communications made by all of their subscribers.20 They would also likely have
had a lot to say about the Austrian courts’ novel requirement that Facebook
monitor its users’ posts in search of any image of Glawischnig-Piesczek.
Presumably that would require Facebook to use facial recognition technology. As
the referring court and the AG frame the case, the facial recognition issue
probably isn’t in scope—which is fortunate, since the issue is massively complex
and the Court has seemingly received no relevant briefs or analysis. The question
of how filtering users’ communications generally affects their privacy, though, is
germane to the case, and should very much be part of the Court’s analysis. The
lack of briefing on it is disturbing.
Competition: If the Court opens the door to filtering orders in this case, it will
inevitably cause other platforms to adjust their legal risk assessments, behavior,
and product design—even if the ruling doesn’t technically apply to them. That’s
consequential for Facebook’s competitors. For giant companies like Google
(where I worked until 2015), those consequences may be tolerable—Google has
already spent at least $100 million21 on filtering technology, and spends another
$100 million22 annually on content review. Smaller companies can’t do that,
though. Both immediate compliance costs and long-term uncertainty about what
measures courts might require have major consequences for smaller companies,
making both technologists and investors hesitant to try competing with today’s
incumbent platforms in the first place.23 This concern was apparently discussed
by the Commission in oral arguments, which is good, and illustrates the value of
20 C-293/12 and C-594-12, Digital Rights Ireland Ltd. v. Ireland, (2014).
21 Paul Sawers, YouTube: We’ve invested $100 million in Content ID and paid over $3 billion to
rightsholders, VentureBeat (November 7, 2018) https://venturebeat.com/2018/11/07/youtube-
weve-invested-100-million-in-content-id-and-paid-over-3-billion-to-rightsholders/.
22 David Shepardson, Google spends hundreds of millions of dollars on content review: letter,
Reuters (May 2, 2019) https://www.reuters.com/article/us-alphabet-google-youtube/google-
spends-hundreds-of-millions-of-dollars-on-content-review-letter-idUSKCN1S81OK.
23 Oxera, The Economic Impact of Safe Harbours on Internet Intermediary Start-Ups, Feb. 2015,
https://www.oxera.com/wp-content/uploads/2018/07/The-economic-impact-of-safe-harbours-
on-Internet-intermediary-start-ups.pdf.pdf; Booz & Co., The Impact of U.S. Copyright
Regulations on Early Stage Investment: A Quantitative Study,
http://www.strategyand.pwc.com/media/uploads/Strategyand-Impact-US-Internet-Copyright-
Regulations-EarlyStage-Investment.pdf.
10interventions to articulate points that neither plaintiffs nor defendants have a
strong interest in raising.
Freedom of Expression and Information: Finally, public interest advocates
could have helped elucidate freedom of expression and information issues.
Facebook briefed the relevant law on this issue extensively and well. But the
interests of the company’s users are ultimately distinct from Facebook’s own, and
warrant different representation. Platforms’ and users’ interests do not align
when platforms face choices between protecting users’ expression and protecting
themselves. Removing lawful speech, or adopting flawed enforcement tools that
will inevitably do so, can be the safest and most cost-effective choice for
platforms. It can help them avoid liability, stave off regulation, please advertisers,
or appease influential critics. Many civil society groups charge that this is
precisely what drove platforms to adopt the poorly understood filters that many
use today.
The Internet is an ecosystem. Cases involving major platforms often have
consequences far beyond the companies themselves. Courts should have the
opportunity to hear from the individuals, journalists, political advocates, and
others who will be affected by consequential decisions like this one. In the
absence of those voices, it is unsurprising that the AG is sanguine about
monitoring injunctions, and willing to assume that technology can block bad
information without also blocking good information. The Court should not repeat
that mistake.
III. Making Facebook Monitor and Filter Users’
Posts
The AG concludes that Austrian courts can order Facebook to filter all of its users’
posts for “identical” expression, and monitor the original user for “equivalent”
expression. In this Section, I will briefly describe the relevant EU policy debate on
11monitoring and filtering, which tees up many issues central to this case. I will
then walk through some concerns with the AG’s Opinion, beginning with issues of
fundamental rights. The fundamental rights impact of the proposed filtering
injunction is hard to assess, because it is not clear exactly what content the
Austrian courts and AG believe should be filtered, what errors can be expected,
and whether those errors can reasonably be remedied. Finally, I will review the
AG’s discussion of the eCommerce Directive – which is difficult to follow, but
appears to say that hosts immunized under the Directive can be compelled to use
software-based filters, yet risk serious liability if employees review the filters’
removal decisions. That is a perverse result from a fundamental rights
perspective, and conflicts with both the Court’s precedent and recommendations
of human rights bodies – as well as practices that many platforms have already
adopted at the urging of EU lawmakers.
A. Policy Backdrop
EU policymakers have now invested several years in a major political debate
about whether, when, and how platforms can be compelled to filter their users’
online expression. Lawmakers in Brussels labored for years to reach political
resolution to this question, even in the comparatively straightforward area of
copyright. The Copyright Directive, which was enacted in a series of hotly
contested votes in 2018 and 2019, ultimately included a filtering requirement.
That outcome prompted street protests and newspaper blackouts, as well as a
judicial challenge arguing that the mandate violates Internet users’ fundamental
rights.24 As opponents pointed out during the political process, filters can’t
understand context.25 That means if text, images, or videos violate the law in one
24 Michelle Kaminsky, EU's Copyright Directive Passes Despite Widespread Protests -- But It's
Not Law Yet, Forbes (March 26, 2019)
https://www.forbes.com/sites/michellekaminsky/2019/03/26/eus-copyright-directive-passes-
despite-widespread-protestsbut-its-not-law-yet/#24b0d6902493; Andrew Liptak, Poland has
filed a complaint against the European Union’s copyright directive, The Verge (May 25, 2019).
https://www.theverge.com/2019/5/25/18639963/poland-european-union-copyright-directive-
filed-complaint-court-of-justice.
25 Julia Reda, When filters fail: These cases show we can’t trust algorithms to clean up the
internet, Julia Reda (September 28, 2017) https://juliareda.eu/2017/09/when-filters-fail/.
12situation, filters will likely also block the same material in lawful uses like parody,
journalism, or scholarship. Famous examples include a lecture by Harvard Law
Professor Lawrence Lessig, which was automatically removed from YouTube
because it used music clips to illustrate a legal point.26
The EU’s political branches are now wrangling with a second monitoring
proposal, in the draft Terrorist Content Regulation.27 The Commission and
Council versions of the Regulation required hosts to use upload filters to prevent
content from reappearing.28 They also encouraged platforms to carry out “human
review” or “human oversight,” meaning that platform employees should check
filters’ work and correct errors. Civil society organizations 29 and three UN human
rights rapporteurs30 responded to this new proposed filtering mandate with
alarm. They pointed out that filters would not be able to tell when extremist
material such as an ISIS recruiting video was reused in contexts such as news
reporting, academic research, or counter-radicalization messaging. As a result,
valuable and lawful expression would likely be blocked. Many critics questioned
whether human review of filters’ automated decisions would be sufficient to
offset these harms, noting examples like YouTube taking down over 100,000
videos maintained by the German non-profit Syrian Archive as evidence of war
crimes and human rights violations.31 The most recent draft of the Terrorist
Content Regulation, from the EU Parliament, eliminated the filtering
requirement.32 But many observers worry that the Parliament’s new rapporteur
26 See Reda supra note 25.
27 EU Parliament, Tackling the dissemination of terrorist contest online, Legislative Resolution
(April 17, 2019) https://www.europarl.europa.eu/doceo/document/TA-8-2019-0421_EN.pdf.
28 EU Commission (September 12, 2018) https://ec.europa.eu/commission/sites/beta-
political/files/soteu2018-preventing-terrorist-content-online-regulation-640_en.pdf Art. 6; EU
Council (December 3, 2018) Art. 6,
https://www.parlament.gv.at/PAKT/EU/XXVI/EU/04/57/EU_45743/imfname_10862334.pdf.
29 Article 19 et al, Joint letter on European Commission regulation on online terrorist content,
(December 6, 2019) https://www.article19.org/resources/joint-letter-on-european-commission-
regulation-on-online-terrorist-content/.
30 Supra note 16.
31 Kate O’Flaherty, YouTube keeps deleting evidence of Syrian chemical weapon attacks, Wired
(June 26, 2018) https://www.wired.co.uk/article/chemical-weapons-in-syria-youtube-algorithm-
delete-video.
32 Supra note 27.
13for the trilogue process, conservative Polish MEP Patryk Jaki, will agree to
reinstate the filtering mandate in the law’s final version.33 If so, that law will
likely also, like the Copyright Directive, be challenged before the CJEU as a
fundamental rights violation.
European political discussion has not focused on the kind of filter at issue in this
case: one that is designed to restrict criticism of a public figure, and works by
blocking text—rather than, as is more common, images or video. To my
knowledge, no legislative body to date has seriously proposed this. (A possible
exception is Singapore’s controversial “fake news” law.34) Both defamation and
speech about public figures are widely considered too context-dependent for even
judges to assess easily, and thus particularly ill-suited for enforcement by
automated tools.
A new and more general political debate about filtering is expected in 2019 and
2020, as EU lawmakers consider a proposed Digital Services Act. That legislative
effort is expected to lead to changes in the core law at issue in this case, the
eCommerce Directive—which, as implemented in Member States’ laws, has
defined platforms’ legal obligations for almost two decades.
Alongside the EU’s political debate is a major discussion of filtering from human
rights experts. Representatives of both the UN and regional human rights
systems around the world have raised grave concerns about relying on private
companies and automated tools to police online expression and information.35
The Council of Europe’s Committee of Ministers, in a 2018 Recommendation,
said
33 Patryk Jaki, European Union, European Conservatives and Reformists Group
Member, http://www.europarl.europa.eu/meps/en/197516/PATRYK_JAKI/assistants?.
34 Singapore fake news law a 'disaster' for freedom of speech, says rights group, The Guardian
(May 9, 2019), https://www.theguardian.com/world/2019/may/09/singapore-fake-news-law-a-
disaster-for-freedom-of-speech-says-rights-group.
35 Special Rapporteur on the promotion and protection of the right to freedom of opinion and
expression, 2018 thematic report to the Human Rights Council on content regulation, (2018),
https://www.ohchr.org/EN/Issues/FreedomOpinion/Pages/ContentRegulation.aspx.
14Due to the current limited ability of automated means to assess
context, intermediaries should carefully assess the human rights
impact of automated content management, and should ensure
human review where appropriate. They should take into account
the risk of an over-restrictive or too lenient approach resulting
from inexact algorithmic systems, and the effect these
algorithms may have on the services that they provide for public
debate. Restrictions of access to identical content should not
prevent the legitimate use of such content in other contexts.36
EU lawmakers, to their great credit, have been attentive to the evolving
fundamental rights guidance on Intermediary Liability and filtering. The EU
Commission’s 2018 Recommendation on tackling illegal content online, for
example, says that if hosting providers choose to rely on “automated means” to
review content, they should provide “effective and appropriate safeguards” such
as human review to ensure that “decisions to remove or disable access to content
considered to be illegal content are accurate and well-founded.”37 This attention
to the guard-rails created by fundamental rights will be essential as lawmakers
revise the eCommerce Directive in the coming years. The Court’s ruling in this
case may be one of their most important inputs.
B. Legal Analysis
The AG’s enumerated conclusions about filtering are:
1. The court can order Facebook to filter every post by every one of its users,
in order to block “identically worded” content.
2. The court can order Facebook to filter “equivalent” content, but only from
the account of the user who put up the original post.
3. Once a court determines that specific content is defamatory, Facebook
must take down equivalent content if it is notified about it. (This part
strikes me as relatively uncontroversial.)
36Supra note 16.
37European Commission, Commission Recommendation on measures to effectively tackle illegal
content online, (March 1, 2018)
https://ec.europa.eu/newsroom/dae/document.cfm?doc_id=50095.
15To arrive at these, he relies on some interim conclusions. Importantly, he
concludes that the Court can order Facebook to use automated, software-based
filters, but that it cannot require “active non-automatic filtering.”38 This seems to
mean that the Court could not order Facebook to have its employees carry out
human review of content flagged by filters. In other words, courts can order
Facebook to monitor its users’ posts—but only in the way that experts and human
rights bodies have warned is likeliest to lead to damaging mistakes. As I will
discuss in Subsection 1, that approach is hard to reconcile with fundamental
rights guarantees. But appropriate fundamental rights analysis in this case is
nearly impossible, given the lack of clarity about what is to be filtered, what
errors are likely, and whether those errors can be corrected.
In Subsection 2 I will discuss the eCommerce Directive, and the AG’s troubling
implication that platforms using “active non-automatic” content review in any
situation may forfeit immunity under Directive Article 14. Assuming that this
refers to human review of user content, his reasoning suggests that existing
filtering efforts, including the ones that both small and large platforms adopted at
the urging of EU political bodies, expose the platforms to massive legal risk—not
only in this case, but for any future claims in areas such as copyright, trademark,
or defamation. This assessment of human review is also problematic for
platforms’ ordinary notice and takedown operations. If the simple act of looking
at content immediately creates liability for platforms, they have strong reasons to
avoid reviewing or moderating user content at all—or, if they do moderate, to err
heavily on the side of taking content down. These conclusions are not required by
the eCommerce Directive, and they place an unnecessary and disproportionate
burden on Internet users’ rights. The Court should not accept them.
38 Par. 61.
161. Fundamental Rights
Both the CJEU39 and the European Court of Human Rights (ECtHR)40 have in
the past disapproved rulings that would have effectively required platforms to
proactively monitor users’ expression. The CJEU specifically rejected an
injunction requiring a social media platform to filter users’ communications in a
copyright case involving the Belgian collecting society SABAM. Both courts
identified serious fundamental rights concerns with laws that would require
platforms to monitor their users. One concern is that inspecting users’
communications can invade their privacy or data protection rights. This privacy
issue is little explored in the case law, and not addressed in the AG’s Opinion.
Another concern is that filters can impinge on users’ rights to receive and impart
information, whether by blocking lawful expression and information41 or
preventing platforms from hosting open forums for discussion in the first place. 42
Recent public discussion about filters has largely focused on their errors and
resulting removal of legal expression and information. But the concern about
preserving open forums for online discussion has become increasingly pressing,
as Internet users have consolidated onto a small handful of powerful private
platforms like Facebook, YouTube, and Twitter. These major platforms, seeking
39 Case C-360/10, SABAM v. Netlog NV, (2012); Case C-70/10, Scarlet Extended SA v. SABAM,
(2011).
40 Magyar Tartalomszolgáltatók Egyesülete (MTE) v. Hungary, App. No. 22947/13, Eur. Ct.
H.R. 135 (2016), Par 82 (strict liability for platform “allowing unfiltered comments” in defamation
case violated Convention Article 10); compare Delfi AS v. Estonia, App. No. 64569/09, Eur. Ct.
H.R. 586 (2015) (strict liability in hate speech case did not violate Article 10); Daphne Keller, New
Intermediary Liability Cases from the European Court of Human Rights: What Will They Mean
in the Real World?, Center for Internet and Society (April 11, 2016)
http://cyberlaw.stanford.edu/blog/2016/04/new-intermediary-liability-cases-european-court-
human-rights-what-will-they-mean-real.
41 Netlog Par. 50 (filtering injunction “could potentially undermine freedom of information, since
that system might not distinguish adequately between unlawful content and lawful content, with
the result that its introduction could lead to the blocking of lawful communications”).
42 MTE Par. 61 (defendant “provided forum for the exercise of expression rights, enabling the
public to impart information and ideas”), Par. 86 (strict liability would threaten “the comment
environment of an Internet portal, for example by impelling it to close the commenting space
altogether”). CJEU precedent has not focused on the existence of open forums as a Charter Article
11 concern, but has noted the threat to open forums’ commercial viability, saying that a filtering
mandate would “result in a serious infringement of the freedom of the hosting service provider to
conduct its business since it would require that hosting service provider to install a complicated,
costly, permanent computer system at its own expense”. Netlog Par. 46.
17to avoid controversy or liability in legal grey areas, have prohibited content
ranging from art to images of breastfeeding to photos of indigenous Amazonian
women in traditional garb.43 Many experts fear that the remaining smaller
platforms that offer more open forums for expression may not remain
economically viable if the law requires expensive filters—or if operators fear that
a court might at any moment impose that requirement.
The fundamental rights case law on filtering doesn’t provide definitive guidance
for assessing any of these risks. But it does stand for the proposition that filters’
real-world consequences matter. Courts and lawmakers can’t just assume that
filtering technology is perfect or adequate to meet the Charter’s and Convention’s
requirements. They must factor in filters’ real capabilities or shortcomings in
order to assess how well filters serve legitimate government aims, what burdens
they place on other fundamental rights, and whether that burden is proportionate
or necessary.
The AG’s Opinion makes this assessment hard. He recommends a filtering
injunction without defining or analyzing three very basic things. First, what
specific content will Facebook be ordered to filter? Second, how effective will the
filters be in distinguishing legal from illegal content? Third, can Facebook or
other future platforms affected by the ruling try to correct filters’ errors using
mechanisms like human review? These questions are essential to assess how
43Par Perrine Signoret, Facebook: la justice se penche sur la censure de, Le Monde (February 1,
2018) https://www.lemonde.fr/pixels/article/2018/02/01/censure-de-l-origine-du-monde-sur-
facebook-une-attaque-contre-la-democratie_5250611_4408996.html; Jessica Reed and Becky
Gardiner, The beautiful breastfeeding images Facebook is missing out on, The Guardian
(February 23, 2012)
https://www.theguardian.com/commentisfree/2012/feb/23/breastfeeding-images-facebook-
missing; Waqas, Brazil will sue Facebook for blocking picture of indigenous woman, HackRead
(April 20, 2015) https://www.hackread.com/facebook-blocking-brazil-indigenous-picture/; see
generally Daphne Keller, The EU's Terrorist Content Regulation: Expanding the Rule of
Platform Terms of Service and Exporting Expression Restrictions from the EU's Most
Conservative Member States, Center for Internet and Society (March 25, 2019)
http://cyberlaw.stanford.edu/blog/2019/03/eus-terrorist-content-regulation-expanding-rule-
platform-terms-service-and-exporting.
18many “dolphins” will be caught in filters’ nets—in other words, how often
Internet users around the world will be prevented from receiving or imparting
lawful information. Neither the AG nor the Court should attempt to resolve the
case without clearer answers.
a. What “Identical” or “Equivalent” Content Is to Be Filtered?
The AG says that Facebook can be ordered to monitor all of its users in order to
block “identical” content, and also monitor the original user’s account for
“equivalent” content. Filtering “identical” content, he suggests, is simple. But it is
hard to tell what content he considers to be “identical.” That imprecision is a
problem. Without knowing what content a filter is supposed to detect, it is hard
to predict its likely accuracy, or how much lawful information and expression it
might block by accident.
i. “Identical” Content
The filter for “identical” content that the AG contemplates seems to be one or
more of the following:
Facebook must block every user from sharing copies of the original post using
Facebook’s “Share” button. This is an odd issue to litigate, because taking down
the user’s post should already, as a matter of Facebook’s basic architecture, have
eliminated any “shared” copies. You don’t need a case like this to compel that
outcome. Apparently, the Austrian government raised the “share” function in its
brief, though, so it may be in play.44
Facebook must block every user from posting the text that the Austrian court
ruled defamatory (“‘lousy traitor of the people’ and/or a ‘corrupt oaf’ and/or a
member of a ‘fascist party’”). As best I can discern, this is what the AG means by
44 Par. 56.
19“identical” content.45 If that’s right, the Opinion raises major concerns for
Facebook users’ information and expression rights. A filter blocking these bare
phrases would prevent friends from calling one another oafs or traitors in jest,
and prevent historians or journalists from writing about actual fascists. It would
also block news coverage of this case—or academic and legal analysis, including
the AG’s own Opinion. These kinds of problems with text filters are nothing new.
Wikipedia’s entry on the Scunthorpe Problem documents notorious and often
comic examples of text filtering failures going back to the 1990s. 46 There are very
serious examples, too. Numerous victims of racially-based harassment have gone
online to bear witness to their experiences, only to be penalized or locked out of
social media for repeating words used by their attackers.47
Facebook must block every user from posting the text that the Austrian court
ruled defamatory (“‘lousy traitor of the people’ and/or a ‘corrupt oaf’ and/or a
member of a ‘fascist party’”) coupled with any photograph of the plaintiff. This
seems to have been what the Austrian appellate court had in mind when it said
Facebook must filter “identical” content.48 From the perspective of defamation
law, enjoining posts with images of the plaintiff makes some sense, since she
would not have a right to stop people using those words about anyone but herself.
45 In theory the “identical content” could be the Austrian user’s entire written post. See supra note
3 for text. The AG does not include that text in his Opinion, though. He writes that he takes
“’identically worded items of information’ to mean … precise manual reproductions of the
information which [the Austrian court] has characterised as illegal” (par. 56) and notes that the
injunction below covered allegations “that the applicant was a ‘lousy traitor of the people’ and/or
a ‘corrupt oaf’ and/or a member of a ‘fascist party’.” (Par. 14).
46 Wikipedia, https://en.wikipedia.org/wiki/Scunthorpe_problem.
47 Tracy Jan and Elizabeth Dwoskin, A white man called her kids the n-word. Facebook stopped
her from sharing it, The Washington Post (July 31, 2017)
https://www.washingtonpost.com/business/economy/for-facebook-erasing-hate-speech-proves-
a-daunting-challenge/2017/07/31/922d9bc6-6e3b-11e7-9c15-177740635e83_story.html. A
recent U.S. case strangely merged this too-common fact pattern with the one at issue in
Glawischnig-Piesczek. A politician’s office posted video documenting protestors’ crude
allegations and suggestion of violence against him – then complained when Twitter removed the
video for violating its policy against threats. Marc Rod, Twitter reverses course, unlocks Mitch
McConnell campaign account and leaves video that violated threats policy,
https://www.cnbc.com/2019/08/09/mitch-mcconnell-campaign-twitter-account-is-
unlocked.html.
48 For the second instance Austrian court, “the reference to ‘identically worded items of
information’ was to publications of photographs of the applicant with the same accompanying
text”. (Par. 56, italics altered).
20But the Supreme Court referral asks only about “identically worded items of
information,” so the question to the CJEU appears to be about text.49 The AG
seems to interpret it that way: his Opinion doesn’t discuss photo filtering or the
distinct concerns it would raise.
From the perspective of fundamental rights, requiring a company like Facebook
to employ facial-recognition-based filters would be radical. At a time when many
privacy advocates want platforms to stop rolling out pervasive biometric
identification, an injunction covering any photo of the plaintiff would make the
technology mandatory—requiring Facebook to run facial scans on people who
have nothing to do with this case.50 Facebook’s billions of users may (or may not)
have consented to that. The other people depicted in their photos almost certainly
haven’t. If that’s what this case were about, we’d expect the referral, briefing, and
AG Opinion to discuss the depicted people’s privacy and data protection rights.
More basically, an order to block content based on plaintiff’s image or identity
would raise fundamental rights issues similar to those the Court considered in
Google Spain and is assessing in the newer “Right to Be Forgotten” cases now.
While the doctrinal basis of data protection and defamation claims differ, they
raise similar tensions between a claimant’s rights to reputation, privacy, or data
protection on the one hand, and other Internet users’ rights to seek and impart
information on the other. As I discuss in my more detailed article about Data
Protection and Intermediary Liability, few courts have considered how the
precedent assessing this balance of rights for search engines might apply to the
very different situation of social media hosts like Facebook.51 We do know,
however, that the Article 29 Working Party’s guidelines under Google Spain
called for search engines to evaluate each individual item of online information
49 Par. 56.
50 Mark Scott and Naomi O’Leary, Facebook’s privacy push stumbles over EU rollout of facial
recognition technology, Politico (April 6, 2018) https://www.politico.eu/article/facebook-facial-
recognition-privacy-data-protection-cambridge-analytica-mark-zuckerberg/.
51 Daphne Keller, The Right Tools: Europe’s Intermediary Liability Laws and the EU 2016
General Data Protection Regulation, 33 Berkeley Tech. L.J. 287 (2018) at 322-327.
21separately before de-indexing, and particularly emphasized the complexity of this
review in cases involving public figures.52
ii. “Equivalent” Content
It is also unclear what “equivalent” content Facebook can be required to monitor
on the original user’s account. The AG suggests that text is “equivalent” if it
“scarcely diverges” from the original, containing perhaps a “typographical error”
or “slightly altered syntax.”53 Filters like that might exist for uses like plagiarism
detection, though I’m unaware of any literature on their efficacy and in practice
they would likely depend heavily on human review. In any case, though, the AG
notes that the Austrian court might mean something else.54 This ambiguity makes
the fundamental rights consequences of this part of the injunction particularly
hard to assess.
Relying on filters to discern the meaning of written human communication, or
even its sentiment, is a dubious proposition. One recent study of “natural
language processing” filters found errors in one out of every four to five takedown
decisions, and noted that the errors increased when speakers used slang,
sarcasm, or languages that tech company employees didn’t speak.55 Another
found that automated detection tools disproportionately mis-identified social
media posts in African American English as “toxic.”56
The Court should not uphold the unprecedented filtering injunction in this case
without clearly understanding and describing what that injunction would require
Facebook to do. Without clarity about what information, exactly, is to be filtered,
it is all but impossible to assess the injunction’s impact on fundamental rights.
52 Article 29 Data Protection Working Party, Guidelines on the Implementation of the Court of
Justice of the European Union Judgment on “Google Spain and Inc. v. Agencia Española de
Protección de Datos (AEPD) and Mario Costeja González” C-131/12, (November 26, 2014)
https://perma.cc/6LPR-TFRL.
53 Par. 67.
54 Id.
55 Center for Democracy and Technology, supra note 11.
56 Sap et al, supra note 11.
22b. How Much of the Filtered Content Will Be Illegal?
The AG appears to posit that as long as filters accurately flag identical content,
nearly all of that content will be illegal. In general, he says, “repetitions of an
infringement actually characterised as illegal… should be characterised in the
same way.”57 That factual proposition is, of course, central to the fundamental
rights question in this case. As described above, it is very much contested. Basic
questions about how often filters accurately identify duplicate content, and how
often that content violates the law, have fueled years of argument in political,
legal, and human rights circles.
Purely as a matter of formal logic, the AG’s assumption would make sense if
“identical” content were truly illegal in every context. In that case, filters’
inability to assess context wouldn’t matter. The court would only need to know
whether filters can accurately identify duplicates as a technical matter.
In principle, the Austrian Supreme Court could have framed the question this
way: “assuming that every identical copy of this information violates Austrian
law, is the filtering injunction permissible?” But that’s not what it did. Its
question states only that the initial copy is illegal, and asks the CJEU to decide
whether Facebook can be ordered to filter “other identically worded items of
information.”58 Nothing in its referral asserts that every copy is illegal.
Some content really is, at least in some jurisdictions, illegal in every possible
context. But the only examples I am aware of (outside of countries like China)
are very extreme and harmful content, specifically child sexual abuse imagery
57 Par. 65. A filter’s effectiveness in identifying unlawful content depends on two things: (1) its
technical accuracy in identifying duplicates, and (2) its ability to assess whether a duplicate is
illegal in a new context. If this case were about detecting duplicate images or videos, or about
using facial recognition technology to detect new pictures of the plaintiff, there would be
important factual questions about (1), the filters’ purely technical accuracy. Those questions are
also important if filters are to block “equivalent” content of any sort, including text. For
“identical” copies of words in text files (as opposed to words that appear in image files, etc.)
though, it is probably safe to assume filters will be technically accurate. In that case the key
question is about (2), legality when the words are used in a new context.
58 Par. 22.
23You can also read
